Model-Based Policy Adaptation for Closed-Loop End-to-end Autonomous Driving

We sincerely thank reviewer 7qBw for their insightful and inspiring feedback. We are glad to know the reviewer recognizes the motivation and community value of our work, the generality of the MPA framework, as well as its strong empirical performance. We provide our response to the questions below.

W1. Clarification on the counterfactual data generation procedure.

We thank the reviewer for raising this important clarification question. We will make clarification on each question as follows.

• W1.1 (Q2) How are the reference trajectories constructed? Are they simply the ground truth states of the ego vehicle? How the reference policy is applied (and on which time steps)?

The reference policy is not the ground truth states of ego vehicle. Instead, they are constructed by adding random noise to the base policy's prediction (UniAD, VAD, LTF). In practice, we conduct geometric transformation by rotating the trajectories with a random angles uniformly sampled from -5 degree to 5 degree, and also scaled the longitude motions with random scales from 0.1 to 2.0.

The the counterfactual data generated by this reference policy consists of over 20 counterfactual actions under the same states, and their reward will be annotated by the underlying states from HUGSIM during training time. Similar techniques are adopted by the PDM controller in [1].

Note that this 'reference policy' is only applied during training time to help collect counterfactual datasets. We will modify the term 'reference policy' to 'behavior policy' in our revised manuscript to avoid further misunderstanding.

• W1.2 How is the reward calculated?

The reward is calculated by the underlying states from HUGSIM simulator. It is a weighted sum of route completion, lateral distance, speed, and the collision penalty. We attach a more detailed description of the reward definition in Appendix Section A.1.

• W1.3 A figure or more precise notation might be helpful. For example, Figure 3 provides a BEV overview where the colors are not defined, the trajectories are not clearly visible, and the procedure is not shown over the different timesteps.

We thank the reviewer's suggestion in our visualization and will revise it to be more clear in our revised manuscripts. Currently, the BEV map we used is the identical as the direct BEV map prediction from our base policies, and the one in Figure 3 is the visualization from UniAD. For every step of rollouts, the BEV map will be updated, and the model will take in the most recent observations to correct the actions. We will add a multi-step demonstration to clearly demonstrate the inference-time pipeline of MPA.

W2. Regarding reorganizing on section 2.1 and 2.2.

We thank the reviewer for giving this valuable suggestion on our paper writing. While compounding errors is widely acknowledged in RL community, prior end-to-end driving methods fail to resolve it due to the lack of high-fidelity counterfactual data generation and robust policy adaptation algorithm. We will shorten the section in our revision.

Q1. Clarification on the claim: "…increasing the planning frequency and shortening the planning horizon cannot necessarily close the gap between open-loop and closed-loop performance, unless some effective feedback guidance is provided to the E2E agents."

We thank the reviewer for the careful reading and thoughtful feedback on our presentation. We mainly draw this claim from Figure 2, where the closed-loop trajectory prediction yields to compounding error as the planning horizon enlarges. We also notice from Figure 2 that the L2 prediction error is non-neglegible even under small prediction horizon (0.5s~1s). This suggests that behavior cloning loss is not a strong supervision signal even for a short-horizon policy. Then we propose to use other guidance signal (such as reward in our counterfactual dataset) to help correcting the prediction online. We will add more elaboration between the analysis on Figure 2 and this claim to make the presentation more coherent.

Q3. Qualitative studies on the LTF's performance gain from MPA.

We thank the reviewers for their thoughtful feedback. Since no images are allowed in the response, we cannot visualize the heatmap of predicted trajectories, but we'll add them in our revised version. Alternatively, in the tables below, we analyze the statistics in the driving behaviors of LTF, UniAD, VAD, and MPA correction effect on them.

• We quantify the diversity in both longitude and lateral directions, for both in-distribution and unseen scenarios. We can see a clear difference in the longitude prediction's standard deviation between LTF and other methods without MPA. The metrics are reported with the future waypoints in 3 seconds with a unit of meter. In both tables, LTF base model is more agile with higher standard deviation compared to UniAD and VAD in the longitude direction, which demonstrates their capability to accelerate/decelerate to avoid collision or getting into other safety-critical conditions.
• We also notice for all the methods, after MPA's correction, their standard deviation in longitude (and most lateral, except for VAD in OOD scenes) predictions will increase. This partly reveals some internal mechanism in how the MPA improves the behavior quality: it gives the base model better agility in both longitude and lateral movements.
• Although MPA benefits LTF, the beneficial effect is not as large as the others. Since we use the same random trajectory transformation during the counterfactual data generation perspective, this may lead to worse coverage in LTF in the whole trajectory space, since the base action in LTF is already quite diversified.

Table R1.1 Standard deviation in in-distribution scenarios. Without MPA, LTF already has higher longitude behavior diversity compared to the other base policies.

Table R1.2 Standard deviation unseen scenarios, Without MPA, LTF already has higher longitude behavior diversity compared to the other base policies.

We also quantify the distribution of output action corrections $\Delta \hat{a} \approx a^*-a_{base}$. In the following table, we quantify the standard deviation of lateral behavior correction from MPA in both in-distribution (ID) and unseen (OOD) scenes. MPA(LTF) has the highest deviation in lateral correction, which leads to more unstable steering actions. This potentially reduces the final performance in driveable area compliance (related to DAC) and distance keeping (related to NC and TTC) compared to MPA (UniAD) and MPA (VAD).

Table R2. MPA Correction Analysis. MPA (LTF) has higher lateral deviation compared to the other two methods.

Regarding compact representation: Given our current empirical results, there is no clear evidence that the intermediate representation of LTF has significant harm in the final performance of MPA. This intermediate representation will be fed to the policy adapter, while the Q-value models have similar capacity as they take in the raw camera observations.

Q4. Discussion on the computation efficiency.

We ran our experiment on 2xNVIDIA RTX A5000 GPUs for 1000 rollout steps, and calculate the average time comsumption over different modules in MPA:

Table R3. Inference overhead analysis on MPA.

Compared to the inference speed of base policy, the overhead of the additional parameters from MPA is small (a total of 0.16s, <25% inference time overhead for UniAD, <50% for VAD and LTF), while boosting up the final performance significantly given a considerably large amount of counterfactual samples and multi-step Q value feedback. We will explore optimizing more efficient (parallel) rollout for MPA in the future.

[1] Dauner, Daniel, et al. "Parting with misconceptions about learning-based vehicle motion planning." Conference on Robot Learning. PMLR, 2023.

We thank Reviewer hzLc for recognizing the well definition and industry-relevance of our work, as well as the effectiveness of our unified MPA framework on the nuScenes closed-loop benchmarks. We answer each question in turn as follows.

W1. Regarding insufficient motivation and details on individual components.

Our MPA composes of three major components: (1) counterfactual data generation, (2) diffusion adapter, and (3) Q value network for inference-time scaling.

1. The counterfactual data generation aims to provide better data coverage by conducting random sampling based on the behavior policy. The 3DGS model for the training scenes will be used to synthesize our counterfactual observations with action and reward annotations, composing the training data for diffusion adapter and Q value networks.
2. The diffusion adapter aims to provide diverse residual action proposals to effectively correct the output actions from pretrained base policy, such as UniAD, VAD, etc.
3. The Q-network aims to evaluate the quality of the corrected trajectory proposals from diffusion adapter by different aspects: driveable area compliance, collision avoidance, route following, and speeding constraints.

W2 & W4. Some claims are not substantiated by evidence or results in the current version.

We clarify the links between our claims and the accompanying results below:

1. "The performance drop in closed-loop evaluation stems from observation mismatch and objective mismatch."

   Evidence: This statement follows directly from the decomposition in Eq. (1), which analytically separates the two sources of error. The objective-mismatch term captures the gap between the training loss and the online reward, while the observation-mismatch term accounts for perceptual discrepancies introduced by the simulator.

2. "…the first mismatch is actually minor in the open-loop evaluation."

   Evidence: Figure 2 supports this point: prediction errors for policies using ground-truth visual inputs and those using 3DGS-rendered inputs are nearly identical, indicating that observation mismatch has negligible impact under open-loop evaluation.

3. "the fidelity of the 3DGS-based simulation in its reconstruction quality."

   Evidence: Figure 2 validates that the reconstructed data has minor impact on the prediction error of end-to-end autonomous driving policies, such as UniAD, VAD, and LTF. We further evaluate the rendering quality of novel view synthesis. We can see that the FID scores of 3DGS rendered images are around 20 when the view pose has small lateral displacements with the ground truth trajectories.

We will revise the manuscript to make the presentation more coherent between the analytical derivation and the experimental validation.

W3. In the preliminary experiment, how is the prediction horizon controlled and how are experiments with 3DGS conducted?

We appreciate the reviewer’s request for clarification.

Prediction horizon. Each base policy produces a sequence of future way-points spanning 3 s. To create the horizons shown in Fig. 2, we truncate that sequence after 1, 2, … H steps and compute the L2 error between the retained way-points and the ground-truth trajectory. Thus the horizon is controlled entirely by how many of the already-predicted way-points we evaluate—no additional forward passes are required.

3DGS experiments. Following the HUGSIM protocol, we train a separate 3D Gaussian-Splatting (3DGS) model for every scene. Ground, static background, and dynamic objects are reconstructed as three disjoint Gaussian sets and composited at render time. In the preliminary evaluation, we run each policy twice: (1) on the original nuScenes camera images and (2) on the images rendered by the corresponding 3DGS model. Then we measure prediction errors with the same horizon-truncation procedure described above.

W5. What is the meaning of $r(s_t, a_t)$ and env.step() in the pseudo code?

We apologize for the insufficient explanation in the current manuscript. Based on our MDP's definition in section 2.1, $r(s_t, a_t)$ is the reward feedback that HUGSIM returns to evaluate the performance based on the current driving states and actions. env.step() is called to update the next vehicle states as well as the corresponding visual observations with HUGSIM's 3DGS rendering engine. The collected counterfactual dataset is composed of (s, a, o, r), including the state action transition pairs, visual observations and the reward labels. They will later be used to train the policy adapter and value model.

W6. Icon explanation in Figure 1 and 3.

We thank the reviewer for pointing out this concern. The icons in Figure 1 and 3 include the driving actions (throttle and steering), synthetic counterfactual dataset, frozen base E2E policies, as well as the trainable diffusion adapter and value networks. We will add a legend in our revised manuscript to help the audience better understand our pipeline details.

W7. Additional visualization and more comprehensive results on different datasets.

Additional visualization. Beyond Figure 4, we offer additional visualizations for all nine evaluated methods across four distinct scenarios in Appendix Figures 11 and 12. Each scenario is illustrated with four representative key frames, covering both safety-critical and non-safety-critical traffic contexts. We will include more qualitative examples in the appendix for our final revised version.

Additional baseline comparison. We add additional baseline results on imitation learning with data aggregation (BC-Safe-DAgger), and model-based RL agent (LAW-Imagine).

• DAgger [1] is an online imitation learning approach that extends our BC-Safe baseline in aggregating new online rollout data. We observe that BC-Safe-DAgger performs better compared to BC-Safe, particularly in the safety-critical scenes. But it still cannot outperform our proposed MPA (reported in Table 1 in the original manuscript).
• LAW [2] follows VAD and incorporates additional latent world model loss as an auxiliary regularization when predicting actions during pretraining. We use the output trajectory from LAW as the base action, and further incorporate imagination rollout in MBRL using the same value functions as MPA for LAW, denoted as LAW-Imagine. We use LAW's checkpoint from their public codebase, and report the results of LAW-Imagine on HUGSIM in Table R1. The results show that with world model augmentation and added imagination rollout, LAW-Imagine can outperform three pretrained base policies -- UniAD, VAD, and LTF, but is still not as good as MPA (reported in Table 1 in the original manuscript) due to the significant drop in the comfort of the driving behaviors.

Table R1.1: Additional in-domain results

Table R1.2: Additional results under unseen scenes

Table R1.3: Additional results under safety-critical scenes

Additional dataset evaluation. Besides additional baselines, we also evaluate MPA’s effectiveness on the Waymo dataset as well as on nuScenes. The Waymo dataset contains only three front-facing camera views, whereas nuScenes provides six. Consequently, we report the zero-shot performance of the nuScenes-trained LTF and MPA (LTF) models, which accept three-camera inputs. We evaluate on the first 40 scenes of the “NeRF on the Road” (NOTR) dataset used in EmerNeRF [3]. The results below demonstrate that MPA generalizes well and delivers consistent performance gains on datasets beyond nuScenes.

Q1. Which specific factors contribute most significantly to the reported HDScore?

As we stated in the paper, the HDScore is computed as:

$$\\text{HDScore}=\\text{RC} \\times \\frac{1}{T}\\sum_{t=0}^T \\Big\\{ \\prod_{m\\in \\{\\text{NC, DAC}\\}} \\text{score}_m \\times \frac{\sum\_{m \\in \\{\\text{TTC, COM}\\}} \\text{weight}_m \\times \\text{score}_m}{\sum\_{m\\in \\{\\text{TTC,COM} \\}} \\text{weight}_m} \\Big\\}_t.$$

Our MPA method prevails baselines particularly in the Route Completion (RC) and Driveable Area Compliance (DAC) metrics, which leads to a higher HDScore. Please kindly check Tables 1 and 2 in our original manuscript for more details.

[1] Ross, Stéphane, et al. "A reduction of imitation learning and structured prediction to no-regret online learning." AISTATS 2011

[2] Li, Yingyan, et al. "Enhancing End-to-End Autonomous Driving with Latent World Model." ICLR 2025.

[3] Yang, Jiawei, et al. "EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision." ICLR 2024

We thank Reviewer 2ZCm for the thoughtful and constructive feedback, and for acknowledging the clarity, practical value, and strong empirical performance of our MPA framework in bridging open-loop and closed-loop end-to-end driving. We address each question as follows.

W1.The counterfactual data generation relies heavily on teacher-forcing, which may not explore sufficient behavioral diversity to handle truly out-of-distribution scenarios.

We thank the reviewer for this insightful question. The key goal in our counterfactual generation is to balance the diversity of the sampled trajectories and the fidelity of the rendered images.

• For the diversity of the data, we add random noise by scaling and rotating the predicted trajectories from the base policy. In this way, we can collect the state and action transition pairs under different future motions in the next few frames. We also conduct an ablation study on the counterfactual rollout steps in Figure 5 in our original manuscript. Longer counterfactual rollout steps bring in better diversity in the dataset. We can see that the performance of MPA increases significantly from step 0 to 3, and then converges around 4 and 5 steps. This suggests the current behavioral diversity in the counterfactual dataset is sufficient to train robust policy adapters and Q value models.
• Meanwhile, for the sake of rendering fidelity, we need to constrain the divergence of the sampled counterfactual trajectories. As visualized in Figure 6 in our supplementary material, the quality of the 3DGS-rendered images degrades (FID gets larger than 50) as the lateral distance to the ground truth trajectories gets too large. To avoid this, we will truncate the samples that violates the distance or reward threshold during the counterfactual data generation phase, as included in Algorithm 1 in our main text.

W2. About the visual fidelity to support the observation mismatch claim.

HUGSIM allows us to script diverse, photorealistic behaviors for dynamic objects, enabling us to synthesize realistic safety-critical interactions and mitigate the scarcity of real-world interaction-rich data.

As shown in Fig. 6 of the supplementary material, HUGSIM renders high-fidelity images when the synthesized viewpoint remains close to the ground-truth path (FID ~20). Visual quality degrades only as the lateral deviation grows: once the offset exceeds ~0.8 m the FID can surpass 50. Such large deviations occasionally arise when the base policies (UniAD, VAD, LTF) drift during closed-loop rollouts.

Accordingly, when producing counterfactual data, we will constrain the lateral displacement, ensuring every rendered observation remains photorealistic and avoids perception noise that could degrade MPA's final performance.

W3 & Q3. Further explanation on the Q-value decomposition into the specific components, as well as the weighted sum of different terms.

We follow the prior reward recipes in [1] to decompose the essential value function terms in autonomous driving. Note that some of these terms are under different scales in our case. For instance, the route completion is 0 to 1 across all timesteps, and the difference between different timesteps will be very small. Distance to the driving lane has a unit of meters (m). The speed is m/s, and collision is a binary term of 0 or 1.

In order to select the value coefficients, we zoom into the empirical distribution of these individual reward term among the training data, then balancing them by timing them approximately by the reciprocol of their mean, i.e. $w_\text{collision} \approx 1/\text{mean}(r_\text{collision})$, forming the final weights coefficients as presented in the hyparameter table in appendix Table 4.

W4 & Q4. Computational costs are not discussed.

We thank the reviewer for raising these valuable points. We ran our experiment on 2xNVIDIA RTX A5000 GPUs for 1000 rollout steps, and report the average time consumption among different modules in MPA:

Table R1. Inference overhead analysis on MPA.

Compared to the inference speed of base policy, the overhead of the additional parameters from MPA is small (a total of 0.16s, <25% inference time overhead for UniAD, <50% for VAD and LTF), while boosting up the final performance significantly given a considerably large amount of counterfactual samples and multi-step Q value feedback. We will explore optimizing more efficient (parallel) rollout for MPA in the future.

W5. Limited number of safety-critical scenes.

We thank the reviewer for highlighting this concern. Generating diverse, safety-critical scenarios remains challenging because real-world crash data are scarce and reliable ego policies for automated stress-testing are still under development. Even without explicitly adversarial scenes, we find that state-of-the-art E2E policies (UniAD, VAD, LTF) frequently fail in closed-loop rollouts. Since the main focus of this paper is to first train a robust end-to-end policy, we manually design 10 safety-critical scenarios in the scenes in Singapore from nuScenes dataset.

The current ten safety-critical scenes we curated from HUGSIM also align with other recent benchmarks such as NeuroNCAP [2], where they evaluate the ego policy in five front scenes and five side scenes from the nuScenes dataset.

In the future, we intend to use MPA as a robust ego policy (i.e. "victim policy") that is ready for a more scalable safety-critical scenario generation pipeline in the end-to-end autonomous driving research.

Q1. Clarification on the qualitative results in Figure 4.

We thank the reviewer for raising clarification on our important qualitative results. In Figure 4, we presented two scenarios in two rows. The three columns show the non-safety-critical scenes (left), VAD rollout results under safety-critical scenes (middle), and MPA-VAD rollout results under safety-critical scenes (right). We only pick 2 key frames (pre-crash, and crash frame) for each scenario due to the space limit, highlighting the different outcomes of VAD (crashing) and MPA-VAD (safe) at the end of the evaluation episode. We further illustrate more comprehensive 4-frame qualitative results in Figure 11 and 12 in our supplementary materials. Please kindly check them if interested.

Q2. According to the original HUGSIM paper, UniAD generally performs better than VAD on HDScore, which is contradictory to the authors results. Could you explain why it is the case?

We believe the difference stems primarily from the evaluation set.

• Scope of scenes. The HUGSIM paper reports results on a small, high-quality subset of nuScenes (one scene in their initial release and eight scenes in their latest). To obtain broader coverage, we trained 3DGS reconstructions for more than 360 nuScenes scenes and evaluated on 290 training and 80 for testing (including safety-critical) cases.
• Rendering fidelity and perception sensitivity. While this larger set offers greater diversity, some of the newly reconstructed scenes could lead to lower perception quality than the hand-selected ones in the HUGSIM release. Compared to UniAD, VAD consumes a vectorized BEV scene representation and is more robust, thus ranking higher on our expanded evaluation set.

Despite these differences, MPA boosts both UniAD and VAD on our extensive benchmark and the original HUGSIM scenes. We will clarify this point in the revision and explore a better balance between scene diversity and rendering fidelity in future work.

[1] Chen, Jianyu, Bodi Yuan, and Masayoshi Tomizuka. "Model-free deep reinforcement learning for urban autonomous driving." IEEE ITSC 2019.

[2] Ljungbergh, William, et al. "Neuroncap: Photorealistic closed-loop safety testing for autonomous driving." ECCV 2024.

We express our gratitude to reviewer s9GR for their feedback and for acknowledging our contribution in closed-loop autonomous driving and clear presentation in the methodology pipeline. We provide our response to the questions as below.

Q1. Regarding the 'incremental recombination of known pieces'.

We thank the reviewer for raising this clarification point regarding our major contribution. While pairing synthetic data with a value head is well-explored in MBRL, our proposed MPA addresses an open challenge in the real world problem: photorealistic end-to-end closed-loop driving. MPA boosts HDScore over UniAD/VAD/LTF in a scalable evaluation pipeline using nuScenes dataset. To our knowledge, no prior MBRL or E2E driving work unifies sensor-level counterfactual generation, diffusion-based residual control, and online multi-principle value guidance in a single real-time system. We believe that MPA could offer a meaningful advance in E2E driving problem beyond an incremental recombination.

Q2 & Q6. The motivation for all the submodules (simulator, diffusion adapter, Q-network) is not clear. The corresponding ablation results (on these submodules) are unclear.

Motivation of each module: We thank the authors for raising this important aspect of our key methodology.

• The adopted simulator HUGSIM provides a systematic benchmark for closed-loop end-to-end autonomous driving, which is missing in the preexisting open-loop evaluation in end-to-end autonomous driving research. The 3DGS model for the training scenes will be used to synthesize our counterfactual observations with action and reward annotations, composing the training dataset of MPA.
• The diffusion adapter aims to provide diverse residual action proposals to effectively correct the output actions from pretrained base policy, such as UniAD, VAD, etc.
• The Q-network aims to evaluate the quality of the corrected trajectory proposals from diffusion adapter by different aspects: driveable area compliance, collision avoidance, route following, and speeding constraints.

Ablation results of each module: The effectiveness of counterfactual data generation is visualized in Figure 5, where more counterfactual rollout steps yield to better final performance. The ablation studies in Table 3 of our original manuscript show the effectiveness of Q value function guidance and the policy adapter.

Q3. What advantage does the diffusion adapter offer?

As introduced in Q2, the diffusion adapter offers more diverse and high-quality proposals of action correction, which are learned based on the behavior from counterfactual data. This effect of the policy adapter is demonstrated in the ablation studies of Table 3 in our original manuscript, where method without policy adapter (ID-5) suffered performance degradation compared to the full model, especially under safety-critical scenarios.

Q4. Discussion and comparison with Model-Based RL approaches.

We thank the reviewer for raising the line of model-based RL research. World model indeed plays an important role in end-to-end autonomous driving. We will add more discussion and references in our revised manuscripts with existing MBRL and works like PlaNet, Dreamer as the reviewer mentioned.

Besides the current baselines, we compare our method with a latent world model-based end-to-end driving framework LAW [1]. LAW follows VAD and incorporates additional latent world model loss as an auxiliary regularization when predicting actions during pretraining. We use the output trajectory from LAW as the base action, and further incorporate imagination rollout in MBRL using the same value functions as MPA for LAW, denoted as LAW-Imagine. We use LAW's checkpoint from their public codebase, and report the results of LAW-Imagine on HUGSIM in Table R1.1-R1.3. The results show that with world model augmentation and added imagination rollout, LAW-Imagine can outperform three pretrained base policies UniAD, VAD, and LTF in route completion, and it prevails in non-collision and time-to-collision. However, the comfort of LAW-Imagine is the far worse than the other approaches, which means the agents possibly driving in a zig-zag manner to guarantee safety. As a result, the Route completion and HDScore of LAW-Imagine is not as good as MPA-enhanced base policies.

Q5. Comparison with imitation learning with online data aggretation.

We extend our BC-Safe baseline [2] with online data aggregation [3], i.e. rolling out new counterfactual dataset based on the trained policy and aggregate them to the next round of policy learning. We conduct one round of such data aggregation on the training dataset, and report the performance of BC-Safe-DAgger for both training (in-domain), testing (unseen scenes), and safety-critical scenes in Table R1.1-R1.3. The results show that data aggregation helps improve BC-Safe's route completion performance for in-domain setting and safety-critical setting, but has limited improvements in the unseen scenes.

Table R1.1: Additional in-domain results and comparison with existing base policies with MPA. Bold means the best for individual metrics.

Table R1.2: Additional results under unseen scenes and comparison with existing base policies with MPA. Bold means the best for individual metrics.

Table R1.3: Additional results under safety-critical scenes and comparison with existing base policies with MPA. Bold means the best for individual metrics.

Q7. Missing literature review on counterfactual data augmentation in Model-Based RL.

We thank the authors for raising these related works, indeed some of the counterfactual data augmentation works contribute to the generalizability of MBRL agents, e.g. [4, 5]. We will add the discussion on them in the related work in our revised manuscript.

[1] Li, Yingyan, et al. "Enhancing End-to-End Autonomous Driving with Latent World Model." ICLR 2025.

[2] Pan, Yunpeng, et al. "Agile autonomous driving using end-to-end deep imitation learning." RSS 2018.

[3] Ross, Stéphane, et al. "A reduction of imitation learning and structured prediction to no-regret online learning." AISTATS 2011.

[4] Pitis, Silviu, Elliot Creager, and Animesh Garg. "Counterfactual data augmentation using locally factored dynamics." NeurIPS 2020.

[5] Pitis, Silviu, et al. "Mocoda: Model-based counterfactual data augmentation." NeurIPS 2022.

Dear Reviewer s9Gr,

Thank you for engaging in our discussion. Since the discussion period is coming to an end, we'd like to see if you have any additional feedback regarding our previous response. Thank you again for your time during the review and discussion period.

Best,

Authors of Submission #26821